Lithography Wars

In July 2002, at a SEMATECH-organized lithography workshop in front of more than two hundred attendees, a soft-spoken Taiwanese researcher walked to the lectern and told the world’s stepper engineers that their roadmap was a fantasy. Burn Lin was sixty years old. He had spent the previous twenty-two years inside IBM’s research laboratories doing some of the foundational thinking that turned visible-light photolithography into a deep-ultraviolet science, and he had recently taken a senior research role at TSMC in Hsinchu. The agenda for that day’s workshop, organized by SEMATECH and ASML, was supposed to be how to extend lithography at a wavelength of 157 nanometers, an exotic ultraviolet number that the entire industry had agreed for the better part of a decade would be the next step beyond 193 nanometers. Lin had been booked to talk about the path forward. Instead, in a phrase that several people in the room would remember years later, he told them the 157-nanometer emperor had no clothes.

The case Lin laid out was almost embarrassingly simple. At 157 nanometers, the only optical material that transmitted light without absorbing most of it was calcium fluoride, a brittle crystal that nobody had figured out how to grow at the size, purity, and birefringence that production lithography required. Photoresists at 157 nanometers worked badly and outgassed in ways that contaminated the lens. The pellicle, the thin protective film over a photomask, did not exist at all in any usable form. The industry had spent, by Lin’s count and SEMATECH’s, on the order of two billion dollars across its supplier base trying to make 157 nanometers work, and it was not working. Lin’s alternative was so ordinary that for a year nobody in industry had taken it seriously. Stay at 193 nanometers, he said, the wavelength of an argon-fluoride excimer laser already in production. Put a thin layer of purified water between the last lens element and the wafer. Water at that wavelength has a refractive index of roughly 1.44, which means the effective wavelength inside the water shrinks to about 134 nanometers. The numerical aperture of the lens system, the figure that ultimately governs resolution, climbs above one for the first time in the history of optical lithography. It bought, Lin argued, several more nodes of Moore’s Law without any of the calcium-fluoride heroics.

The idea was not new. Lin himself had described it in print in the 1980s while he was still at IBM, and microscope makers had used immersion oil for over a century. What was new was the willingness to take it seriously as a manufacturing path. By the time the SEMATECH workshop ended, the room was divided. ASML’s chief lithography researcher, Jan Mulkens, who had been quietly thinking about the same idea since attending an industry conference in December 2001, returned to Veldhoven and began assembling a small team to test water under a 193-nanometer lens. Nikon’s representatives, who would have to bet a multi-billion-yen investment program against Lin’s argument, went home and waited. Inside Intel’s process organization, a different kind of conversation began. Within six months, the world’s largest chip company had quietly concluded that 157 nanometers was finished. In May 2003, Intel made it public. The 157-nanometer line item in the supply chain collapsed within a quarter.

To understand why that one room mattered as much as it did, the story has to back up nearly a quarter century, to the moment Japan first took stepper lithography away from the Americans.

By 1980, when Nikon shipped its first commercial wafer stepper, the NSR-1010G, to NEC and Toshiba, the company had already been making precision optics for sixty-three years. Its first lenses were ground for the Imperial Japanese Navy in 1917 to keep the navy’s binoculars from depending on Karl Zeiss in Jena. The shift from naval optics to camera lenses to microscope lenses to wafer steppers had, by Japanese industrial standards, been a steady walk down a single road. Nikon’s stepper engineers, working under the umbrella of MITI’s VLSI Project consortium, had bought a GCA stepper through normal commercial channels, taken it apart in a Tokyo laboratory, and built a Japanese answer that improved on it in throughput and overlay accuracy. By 1984 Nikon’s market share matched GCA’s; by 1985 it surpassed it; by 1989 the Japanese, with Canon following Nikon’s playbook two beats behind, held about seventy percent of the global lithography-equipment market.

Canon’s path was less elegant and almost as effective. The company had been making cameras and copiers since the 1930s, and its lithography division began in 1970 with a contact aligner. Through the 1980s it pushed steppers into the second tier of Japanese fabs, and by the second half of the decade it had a substantial position in DRAM lithography, particularly at customers that wanted a hedge against single-vendor dependency on Nikon. Canon’s lenses were never as sharp as Nikon’s. Canon’s stages were never as accurate. But the company’s manufacturing discipline was first-rate, its prices were aggressive, and it understood the politics of multi-source procurement at customers like Toshiba and Hitachi better than anyone outside Japan. Through the 1990s, Canon and Nikon together held a duopoly that, by 1995, accounted for about 78 percent of the world stepper market. Nikon held roughly 49 percent. Canon held roughly 29 percent. The third name on the list, with about 16 percent, was a Dutch joint venture that almost no one in Tokyo took seriously.

ASML had been founded on the first of April, 1984, in a leaky shed at Philips’s Strijp campus in Eindhoven. Its mandate was a single product, the PAS 2000 stepper, that Philips had developed for its own internal fabs and concluded was too far from its core consumer-electronics business to keep. The fifty-fifty joint venture with Advanced Semiconductor Materials International gave Philips a way to walk the technology out without writing it off. Forty-seven of the roughly fifty Philips engineers from the original wafer-stepper group transferred over. The first chief executive, Gjalt Smit, a Dutch astrophysicist by training, had taken the job partly because he had been told the situation was hopeless and partly because he had run out of better offers. The PAS 2000 used hydraulic oil to drive its wafer stage, and by every customer report the oil leaked. One early employee remembered that after a clean-up the customer demonstration room “smelled like an old car shop.”

For the first three years the company operated at a loss, sometimes a deep one. ASMI sold its half of the business back to Philips in 1988 because it could not afford the capital calls. The PAS 2000 went through a redesign that replaced the hydraulics with electric linear motors, a switch made possible by Philips’s licensing of its in-house motor patents to its struggling spinout for nine hundred and thirty thousand dollars. The redesigned PAS 2500 found a small but loyal customer base. None of it was profitable. According to ASML’s own retrospectives, the supervisory board considered shuttering the operation more than once. The decision that kept the company alive was a thirty-million-guilder investment, championed inside Philips by a board member named Henk Bodt over the objections of colleagues who wanted to cut the losses. The money funded the engineering of what would become the PAS 5500. When that machine shipped in 1991, ASML had its first commercially viable platform.

The PAS 5500 was not a technical revolution. It was, in retrospect, a manufacturing one. The architecture was modular in a way that Nikon’s and Canon’s monolithic systems were not: the optical column, the stage, the alignment system, the reticle handling, the resist track interface, all came apart at well-defined seams and could be replaced or upgraded without rebuilding the rest of the machine. A field-service team could pull a failed component from a customer fab, slot in a replacement, and have the tool back in production within hours. Nikon’s machines, hand-tuned to a higher initial standard, took longer to repair and longer to upgrade. As process nodes shrank from 350 nanometers to 250 to 180, the PAS 5500 carried its customers through three or four generations on the same chassis. IBM, hunting through its supplier base in the mid-1990s for a stepper vendor that could keep up with its 0.25-micron roadmap, ordered PAS 5500s over the Japanese alternatives. So did Micron. So did a young Taiwanese foundry named TSMC, building out its first 200-millimeter fabs in Hsinchu and looking for a tool partner that would treat it as a peer rather than a second-tier customer.

By the late 1990s, ASML had a viable business and roughly a quarter of the market. The company’s chief executive through that decade, Willem Maris, who had taken over in 1990 and would hand off in 1999, ran ASML the way an engineer runs a hard problem: by spending most of his time on the factory floor and at customers’ fabs, without much patience for hierarchy. The machines kept getting better. The customers kept buying them. Nobody in Tokyo treated the Dutch as a strategic threat, because nobody in Tokyo had to. Through the 1990s, Nikon and Canon were still the leaders. ASML was the third.

What changed the trajectory was a wager on a different physical architecture, made just before the millennium and shipped in 2001. The internal name was TWINSCAN. The idea, championed inside ASML by a precision-mechanics engineer named Bert van der Pasch, was to put two wafer stages inside a single tool instead of one. While one stage carried a wafer through the exposure column under the lens, the second stage carried the next wafer through a separate metrology station that measured its surface, mapped its alignment marks, and prepared it for exposure. When the first wafer finished, the two stages swapped positions in a single mechanical motion, and the prepared second wafer began exposure within seconds. The first wafer rolled out for unloading; a third wafer rolled in for measurement. The lens, the most expensive subsystem in the tool, never sat idle.

The productivity arithmetic was unforgiving for any single-stage competitor. A traditional stepper spent perhaps a third of its cycle on alignment and metrology and the remainder on exposure. TWINSCAN, by parallelizing the two activities on different wafers, recovered that third. On a tool that listed for thirty million dollars and produced perhaps a hundred wafers an hour, a thirty-percent throughput improvement translated to tens of millions of dollars per year of additional output for the customer. The first TWINSCAN, the AT:750T, shipped to TSMC in August 2001 at the 130-nanometer node. By 2003, every leading-edge fab in the world had at least one. Nikon, whose engineers had considered and rejected a dual-stage architecture as mechanically too risky, had no answer.

The TWINSCAN bet might have been enough on its own to put ASML ahead. Coming six months before Burn Lin’s SEMATECH speech, it set up the second bet that decided the lithography war for a generation.

Inside ASML’s headquarters in Veldhoven, the immersion-lithography idea moved from physics-experiment to development program with unusual speed. Mulkens’s team had begun running water under a test column in early 2002. By autumn 2003 they had a prototype, designated the AT:1150i, that imaged sub-90-nanometer features through a millimeter of moving water held under the last lens element by a hydrodynamic seal designed in part with help from the fluid-dynamics group at Twente, including the physicist Detlef Lohse. Carl Zeiss, ASML’s lens partner since the late 1980s and the supplier whose Jena and Oberkochen workshops had been making the strategic difference between ASML and the Japanese for fifteen years, redesigned the optical column to tolerate the water meniscus. In December 2003 ASML announced the TWINSCAN XT:1250i, the first immersion lithography tool offered for production. TSMC had already committed to buy one, in what was effectively a continuation of Burn Lin’s argument from the previous summer. The first system shipped to Hsinchu in September 2004. IBM took delivery of one in parallel and ran its own evaluation in East Fishkill. By December 2004, both fabs had separately produced functional 90-nanometer chips with one mask layer printed through immersion. The technology worked.

Nikon, watching this from Tokyo, made the decision that broke its dominance.

The internal Nikon argument, reconstructed later from interviews with several of its lithographers, ran as follows. The 157-nanometer dry path was difficult but not impossible. Nikon had invested heavily in calcium-fluoride lens engineering and in 157-nanometer scanner prototypes, and to abandon that program in 2003 would mean writing off a substantial fraction of the company’s research-and-development spend for the prior five years. Immersion at 193 nanometers was, in the Nikon view, an interim solution that would buy two or maybe three nodes before the industry had to confront a wavelength change anyway, at which point Nikon’s 157-nanometer assets would matter again. The water meniscus introduced manufacturability questions that Nikon’s senior engineers were not, in 2003, willing to bet a flagship product on. Better, they concluded, to develop their own immersion path on a slower timeline, hold the existing dry-193 customer base, and pick up the next inflection through 157 if it came back.

The decision was rational on Nikon’s own terms. It was wrong about the world. The 157-nanometer path did not come back. The 193-nanometer immersion path proved manufacturable faster than even ASML had hoped. By the time Nikon’s first immersion prototype shipped in October 2004, ASML was already in a customer’s fab. The follow-on tools, the XT:1700i and XT:1900i, lifted the numerical aperture above 1.2 and then above 1.35, pushing optical lithography to resolutions that would, with subsequent multi-patterning techniques, take chipmakers all the way down to seven nanometers. In the immersion-tool market that opened in 2006, ASML took a 72-percent share against Nikon’s 28 percent in the first year alone. By 2011, ASML’s immersion share was 82 percent against Nikon’s 18, and Canon, which had elected not to enter the immersion market at all, had begun to retreat to legacy applications and announced a long bet on a different technology entirely.

The headline market numbers tell the same story in slower motion. In 2001, the year of the first TWINSCAN, ASML held 22 percent of the global lithography market by units; Nikon held 42 percent; Canon held 35 percent. By 2009, ASML held 67 percent. By 2015, the figure was above 80 percent by revenue. Nikon’s lithography business, which had made the company the world’s most important supplier of stepper tools through the 1980s and 1990s, became a steadily shrinking line item, profitable only at the trailing edge of nodes that the leading-edge fabs had moved past. By the mid-2010s Nikon’s lithography revenue was a fraction of its camera revenue, and the camera market itself was contracting under pressure from smartphones. The strategic monopoly Japan had taken from the Americans in the 1980s, won through patient optical engineering and government-backed industrial coordination, had been transferred to a Dutch company that, twenty years earlier, almost no one in the industry had taken seriously.

Inside ASML, the personalities who carried the company through this stretch were unfussy by the standards of a billion-dollar global business. Doug Dunn, the British executive who had succeeded Maris in 1999 and presided over the TWINSCAN launch, was a former Plessey and Philips man with a flat managerial style. Martin van den Brink, who joined ASML in 1984 as a young engineer, became chief technology officer in the early 2000s, and was eventually called inside the company simply “Mr. ASML,” handled the technical bets. Eric Meurice, who took the chief executive’s chair from Dunn in 2004, drove the immersion ramp. The arrangement that held all of it together was less the structure of any single company than a tightly wound network of supplier relationships: Zeiss in Oberkochen for optics, Trumpf in Ditzingen for laser sources, Cymer in San Diego for excimer lasers, the Eindhoven and Twente engineering schools for talent, ASML’s own customers from TSMC and Samsung to Intel and IBM acting as effectively unpaid co-developers of every new platform. By the late 2000s, that network of relationships had become as hard to copy as any of ASML’s machines.

The pattern that had played out at GCA in the 1980s, with quality and supply discipline rather than headline technology deciding who survived, played out a second time, with the directions reversed. Nikon and Canon had won the 1980s by being the more disciplined manufacturers against an American incumbent that was busy congratulating itself. ASML won the 2000s by being the more disciplined platform-builder against Japanese incumbents that, by then, had become the comfortable ones. The moment that decided the war was not a research breakthrough or a board-room reorganization. It was a customer’s chief researcher walking to the front of a workshop, looking at a roomful of engineers, and saying that the path everyone had agreed to take was an illusion. Burn Lin’s speech in July 2002 lasted maybe an hour. ASML built a monopoly out of believing him.

By 2010, the only question left in leading-edge optical lithography was how many more nodes water and ArF could carry. The wavelength would not shrink again until something entirely new arrived from a different direction, out of a research program that had been quietly grinding through national laboratories and a Veldhoven side project for almost two decades. The Dutch, having taken Japan’s monopoly, were already preparing to defend it.